Hi, I’m having issues with Drama 3 Offload Studio on Windows with an NVIDIA RTX 3090 (24 GB VRAM).
The full model fails during startup with this error:
RuntimeError: NVIDIA GeForce RTX 3090 (compute capability 8.6) has no FP8 tensor cores; Drama 3's FP8 experts need an RTX 40-series / Ada, Hopper or Blackwell GPU (8.9+)
I then installed and launched Small GGUF. It loads successfully and the Web UI opens, but audio generation stalls or stops early. Here are the relevant logs:
[small] drama3-small loaded from GGUF in 88s: 9.10 GiB on cuda:0
[engine] NVIDIA GeForce RTX 3090, 82 SMs, 24 GiB, torch 2.13.0+cu130: one decode step 335.64 ms on the GPU; 9.9 GiB peak
[pause] stall: a drone from frame 23; the take ends there
[space] take 0AewuUUeTrOC: silent or incomplete opening, retrying once
[pause] stall: a drone from frame 23; the take ends there
Later takes produced only about 2.8–3.1 seconds of audio, taking around 42 seconds to generate, and ended with “stall” or “timeout”.
Is this expected on an RTX 3090, or is there a configuration or backend fix for Small GGUF? Also, could the full-model menu option mention the FP8 GPU requirement alongside the 24 GB VRAM requirement?
Hi, I’m having issues with Drama 3 Offload Studio on Windows with an NVIDIA RTX 3090 (24 GB VRAM).
The full model fails during startup with this error:
RuntimeError: NVIDIA GeForce RTX 3090 (compute capability 8.6) has no FP8 tensor cores; Drama 3's FP8 experts need an RTX 40-series / Ada, Hopper or Blackwell GPU (8.9+)
I then installed and launched Small GGUF. It loads successfully and the Web UI opens, but audio generation stalls or stops early. Here are the relevant logs:
[small] drama3-small loaded from GGUF in 88s: 9.10 GiB on cuda:0
[engine] NVIDIA GeForce RTX 3090, 82 SMs, 24 GiB, torch 2.13.0+cu130: one decode step 335.64 ms on the GPU; 9.9 GiB peak
[pause] stall: a drone from frame 23; the take ends there
[space] take 0AewuUUeTrOC: silent or incomplete opening, retrying once
[pause] stall: a drone from frame 23; the take ends there
Later takes produced only about 2.8–3.1 seconds of audio, taking around 42 seconds to generate, and ended with “stall” or “timeout”.
Is this expected on an RTX 3090, or is there a configuration or backend fix for Small GGUF? Also, could the full-model menu option mention the FP8 GPU requirement alongside the 24 GB VRAM requirement?